Weakly supervised semantic segmentation (WSSS) often struggles with classifiers exploiting spurious correlations between objects and backgrounds, leading to poor generalization. We propose shortcut mitigating augmentation (SMA), which generates synthetic object-background combinations by disentangling their features. This reduces reliance on shortcut features and helps classifiers focus on the target object. We analyze the classifier\\\\\\\'s shortcut behavior using an attribution method-based metric and achieve improved segmentation performance on PASCAL VOC 2012 and MS COCO 2014.